Comment by atomicnumber3
1 year ago
It's weird to me that you group LLMs in with code analysis based LSP autocompleters.
Normal autocomplete is basically a context-aware AST browser. It lets you navigate the program semantically instead of file-based.
LLMs are token predictors.
The parent did the same, but of course they're different. The spirit of the question seemed to bundle LSP/autocomplete/LLMs under tooling that some do without.
The way I use LLMs a lot of the time is to simply let copilot complete the method parameters that I was already planning on typing or autocompleting the fields of a struct. It's barely a step over LSP/autocomplete tools but because LLMs have access to more context and aren't restricted by the things they can autocomplete, they work in more scenarios. LLMs get things wrong occasionally but it's usually easy to catch and LSPs can fix those errors.